KarstenSamaschke
Back to all writing

Field notes

I do not particularly like Meta, but this is the right move

Meta put Muse Glimmer out with open weights you can actually run, and Spark 1.2 is still a promise until it lands. I appreciate and support this, despite it being Meta.

Karsten Samaschke · · 4 min read

I do not particularly like Meta, but this is the right move

I do not particularly like Meta, but I still think this is the right move.

They put Muse Glimmer out with open weights under Apache 2.0, small enough that you can run it on a laptop, and Muse Spark 1.2 weights are supposed to follow, which I will only believe when it actually lands.

I appreciate and support this, despite it being Meta, because OpenAI and Anthropic are currently playing the scaremonger game, so regulation slows everyone else down, especially open weights, and then Meta of all companies hands you weights you can actually run in an environment you control. I am not naive about why they do it, and I still prefer that over a closed model with a safety brochure.

What actually landed

Muse Glimmer is out. It is a 30 billion parameter open-weight model from Meta Superintelligence Labs, released under Apache 2.0, and the files are on Hugging Face. It is meant to run locally, on a Mac or a PC with a single consumer GPU, for agent-style work rather than another hosted chat box.

Muse Spark 1.2 is a different story. Zuckerberg said the weights would be opened, and that is still a promise. I will only believe it when it actually lands.

That distinction matters in practice, because a press release is not a checkpoint, and a checkpoint is not yours until you can download it, run it, and keep it.

Why I support it anyway

I am not suddenly a Meta fan. I am not naive about their motives either. They have a business to run, they have an investor story to tell, and they have spent years proving that open weights can be a distribution strategy as much as a gift.

I still support this, because the thing I care about is not the logo on the crate. The thing I care about is whether I can run the model in an environment I control.

OpenAI and Anthropic are currently playing the scaremonger game, so regulation slows everyone else down, especially open weights and cheap Chinese models, and on top of that the stock market is supposed to buy how dangerous, and therefore how valuable, the closed thing is. Meta, of all companies, then hands you weights. That is a strange week, and it is still the right move.

Sovereignty is not a vendor swap

Swapping OpenAI for Mistral, or OpenAI for Meta, is the wrong emphasis. You cannot damn someone for running their own hardware in an environment they control. That is the correct instinct, even when the weights come from a company you do not particularly like.

Sovereignty does not require everything to sit in your own data center all the time. That is better if it is the case, but it is not a must-have. The decisive question is whether you can take the files, run them, and not depend on someone else's API remaining available, remaining cheap, and remaining willing to answer.

Glimmer is small enough that this is not a theoretical speech. You can actually try it.

What I will not do

I will not write a hymn to Zuckerberg. I will not pretend Spark 1.2 is open until the files are there. I will not treat Apache 2.0 as a personality trait of Meta.

I will take the weights that landed, run them where I decide, and keep an eye on whether Spark 1.2 actually shows up.

Sources used